Integration of intraoperative data in interpretable machine learning models to predict postoperative AKI in noncardiac surgery patients

Q: What are the key findings of this study?

A gradient boosted tree model combining preoperative and intraoperative data accurately predicted stage 2/3 postoperative AKI, adding minimal benefit over preoperative data alone.

Q: What is the clinical evidence from this study?

Study design: Cohort. Population: Postoperative acute kidney injury (n=46204). Intervention: Machine learning models using combined preoperative and intraoperative data vs. Models using preoperative data alone or clinical baselines. Primary outcome: Discrimination (AUROC) for stage 2/3 postoperative AKI (AUROC 0.896, 95% CI 0.878-0.914).

Q: What does this research mean for the field?

Machine learning models accurately predict stage 2/3 postoperative acute kidney injury using preoperative data alone, with intraoperative data offering minimal incremental discriminative benefit. Novelty: ClaimNovelty.CONFIRMATORY. Consensus alignment: ConsensusAlignment.SUPPORTS_CONSENSUS.

Justin Do; Karan H. Shah; Melissa Chen Xu; Andrew Hyunwoo Kim; Vivaswat Suresh; Nidhir Guggilla; Michael Li; Rishi Kothari

doi:10.1093/jamiaopen/ooag092

May 5, 2026JAMIA OpenOpen Access

Integration of intraoperative data in interpretable machine learning models to predict postoperative AKI in noncardiac surgery patients

View Full Paper

Ask AI

Bookmark

Share

Key result

Machine learning predicts postoperative AKI well but intraoperative data adds minimal benefit over preoperative variables.

AUROC 0.896
95% CI 0.878-0.914
n=46,204

Figures

Why the study?

The study aimed to determine changes in discrimination when adding intraoperative data to preoperative data and compare tabular machine learning with a time-aware LSTM-based model for predicting postoperative acute kidney injury.

Does the addition of intraoperative time-series data improve the performance of machine learning models in predicting stage 2/3 postoperative AKI in noncardiac surgery patients compared to models using preoperative data alone?

Key Points

To assess discrimination changes when adding intraoperative data to preoperative features for predicting postoperative stage 2/3 acute kidney injury and compare tabular machine learning models with a time-aware LSTM architecture.
Evaluated a retrospective cohort of 46 204 adult patients undergoing 57 055 eligible noncardiac surgeries within the INSPIRE database.
Extracted 38 preoperative and 49 intraoperative variables, defining the primary outcome as KDIGO stage 2/3 postoperative acute kidney injury.
Trained tabular machine learning models using summary statistical features alongside a hybrid MLP+LSTM model and compared them against ASA-PS ≥3 and adapted GS-AKI risk baselines.
Combined-feature gradient boosting trees achieved the highest AUROC at 0.896 (95% CI, 0.878-0.914), closely followed by combined AutoGluon (AUROC 0.893, 95% CI, 0.877-0.909) and preoperative-only gradient boosting trees (AUROC 0.891, 95% CI, 0.871-0.910).
Machine learning models substantially outperformed clinical baselines, with ASA-PS ≥3 yielding an AUROC of 0.723 (95% CI, 0.700-0.746) and adapted GS-AKI yielding an AUROC of 0.719 (95% CI, 0.700-0.739).
The time-aware hybrid MLP+LSTM architecture reached an AUROC of 0.870 (95% CI, 0.848-0.892), failing to outperform simpler tabular machine learning models.

Study Design

Type

Cohort (n=46,204)

Multicenter

No

Structured PICO

Does the addition of intraoperative time-series data improve the performance of machine learning models in predicting stage 2/3 postoperative AKI in noncardiac surgery patients compared to models using preoperative data alone?

P

Population

46,204 adults undergoing noncardiac surgery evaluated to predict stage 2/3 postoperative acute kidney injury using machine learning models on perioperative data.

E

Exposure

Machine learning predictive models (Gradient Boosted Tree, AutoGluon, Multilayer Perceptron + Long Short-Term Memory hybrid) trained on a combination of 38 preoperative variables and 49 intraoperative variables (physiologic monitoring and transfused products sampled every 5-10 minutes).

C

Comparator

Machine learning models trained on preoperative tabular data alone, and traditional clinical baseline comparators (ASA-PS ≥ 3 and adapted General Surgery Acute Kidney Injury [GS-AKI] risk index).

O

Outcome

Stage 2 or Stage 3 postoperative acute kidney injury (AKI) defined by KDIGO serum creatinine criteria (comparing nearest preoperative sCr with the highest postoperative sCr within 48 hours or 7 days post-surgery, or postoperative dialysis).hard clinical

Main Result

Effect estimate: AUROC 0.896 (95% CI 0.878-0.914)

Absolute Event Rate: 0.896% vs 0.891%

Preoperative tabular machine learning models provide highly accurate predictions of severe postoperative acute kidney injury, while adding currently available low-frequency intraoperative data offers limited additional clinical benefit.

Limitations

Data derived from a single South Korean academic center, limiting generalizability.
Low incidence of stage 2/3 AKI constrained the sample size.
Heterogeneity in data reporting prevented external validation against independent cohorts.
Retrospective design subject to missing data and administrative proxies.
Model comparison was not compute-matched, as deep-learning models received a substantially larger tuning and training budget than tabular models like AutoGluon.
Low temporal resolution (<1 Hz) of most intraoperative data may have prevented time-aware architectures from extracting useful physiologic information.

Ask about this paper

Answers are grounded in this paper and cite their sources.

Implication

Preoperative ML models nearly match combined performance for postoperative AKI; leaves open incremental value of intraoperative data in other settings.

Abstract

Objectives We aimed to (1) quantify changes in discrimination when adding intraoperative data to preoperative data and (2) compare tabular machine learning with feature engineering against a time-aware LSTM-based model. Materials and Methods Retrospective cohort of 46 204 adults undergoing 57 055 eligible noncardiac surgery in the INSPIRE database. We extracted 38 preoperative and 49 intraoperative variables; acute kidney injury (AKI) was defined by KDIGO serum creatinine criteria and modeled as stage 2/3 postoperative AKI. Models were trained on preoperative-only and combined pre- and intraoperative data. Intraoperative series were summarized using eight statistical features for tabular models or integrated directly using an MLP+LSTM architecture. Results GBT with combined features achieved the highest AUROC (0.896, 95% CI, 0.878-0.914), followed by combined AutoGluon (0.893, 95% CI, 0.877-0.909) and preoperative-only GBT (0.891, 95% CI, 0.871-0.910). ASA-PS ≥3 (AUROC 0.723, 95% CI, 0.700-0.746) and adapted GS-AKI (AUROC 0.719, 95% CI, 0.700-0.739) underperformed machine-learning models. The hybrid MLP+LSTM model did not outperform simpler tabular models (AUROC 0.870, 95% CI, 0.848-0.892). Discussion The small gain from adding low-frequency intraoperative summaries suggests that most discriminative information for stage 2/3 postoperative AKI was available before surgery. Conclusion Preoperative tabular ML models provided excellent prediction of postoperative AKI, and added limited incremental discrimination at the available sampling frequency. Future work should evaluate whether higher-frequency intraoperative signals better leverage time-aware architectures.

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

Do et al. (2026) conducted a cohort in Postoperative acute kidney injury (n=46,204). Machine learning models using combined preoperative and intraoperative data vs. Models using preoperative data alone or clinical baselines was evaluated on Discrimination (AUROC) for stage 2/3 postoperative AKI (AUROC 0.896, 95% CI 0.878-0.914). A gradient boosted tree model combining preoperative and intraoperative data accurately predicted stage 2/3 postoperative AKI, adding minimal benefit over preoperative data alone.

synapsesocial.com/papers/6a37ae828ae85f346c694609 https://doi.org/10.1093/jamiaopen/ooag092

View Full Paper

Ask AI

Bookmark

Share

Integration of intraoperative data in interpretable machine learning models to predict postoperative AKI in noncardiac surgery patients

Figures

Key Points

Study Design

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Integration of intraoperative data in interpretable machine learning models to predict postoperative AKI in noncardiac surgery patients

Figures

Key Points

Study Design

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading